--- title: "Qwen 4 27B: Release Date, What’s Confirmed, and the Hardware to Prepare (2026)" slug: qwen-4-27b-release-date-specs-hardware-what-is-known-2026 description: "Alibaba named a Qwen 4 27B open-weights model at Apsara on September 22, with no specs, license, or date. What’s confirmed, what the Qwen3.8-27B tells you about it, and which GPU to have ready." author: "Yotta Labs" date: 2026-09-23 categories: ["Inference"] canonical: https://www.yottalabs.ai/post/qwen-4-27b-release-date-specs-hardware-what-is-known-2026 --- # Qwen 4 27B: Release Date, What’s Confirmed, and the Hardware to Prepare (2026) ![](https://cdn.sanity.io/images/wy75wyma/production/51b4e9f68c099daed9abf57a9d1c6423d26fd519-2240x1260.png) *Qwen 4 27B is announced, not released. Here’s what Alibaba actually said, what the current 27B tells you about the next one, and how to be ready the day the weights land.* Of the four Qwen 4 tiers Alibaba named at the Apsara Conference on September 22, the 27B is the one most people are searching for, and for a practical reason: it’s the one you’ll run yourself. Max, Plus, and Flash are API models. The 27B is the open-weights tier for local deployment, the successor to Qwen3.8-27B, which is currently the best model that fits on a single GPU. This post keeps what’s confirmed separate from what’s inferred, and it will be updated the day the model ships. ## TL;DR - Alibaba announced Qwen 4 27B on September 22 as the open-weights, local-deployment tier of the Qwen 4 family; no specs, license, benchmarks, price, or release date were given - Qwen 4 as a whole is in training and coming “very soon” per Alibaba’s Qwen lead; that’s the only timing on record - The precedent is Qwen3.8-27B: a 27B dense vision-language model under Apache 2.0, about 54 GB at BF16, 27 GB at FP8, 16 GB at 4-bit, one GPU - The architecture question is open: the Qwen 4 preview model uses a lean MoE with an n-gram memory table, and whether the 27B stays dense or adopts that design changes the memory math - If you have hardware that runs Qwen3.8-27B today, you’re probably ready; the same 24 GB, 48 GB, and 80 GB tiers are the planning baseline ## Qwen 4 27B status at a glance ## What Alibaba has confirmed At Apsara on September 22, Alibaba announced Qwen 4 as a family of four: Max as the flagship, Plus as the balanced multimodal tier, Flash for low-latency, high-volume work, and a 27B model released as open weights for local deployment. The same day, Qwen project lead Liu Dayiheng said Qwen 4 is in training on a new-generation architecture and will be released “very soon,” with Qwen 4.5 and Qwen 5 to follow at 5 to 10 trillion parameters. That’s the whole official record: a name, a role, and “very soon.” No model card, no parameter count beyond the name, no license, no date. Everything below is inference from the current 27B, from the Qwen 4 preview model, and from Alibaba’s release cadence, and it’s labeled as such. The [Qwen 4 tracking page](https://www.yottalabs.ai/post/qwen-4-release-date-what-is-known-how-to-prepare-2026) covers the family as a whole. ## What the current 27B tells you Alibaba has shipped a 27B dense model in each recent generation, and the line has been the community’s default local model since the Qwen 3.6 generation. [Qwen3.8-27B](https://www.yottalabs.ai/post/qwen-3-8-27b-specs-hardware-requirements-how-to-run-2026), released August 13 to 14, is the current one: 27 billion dense parameters, a vision encoder for image and video input, 262K native context, and an Apache 2.0 license. It scored the same on Artificial Analysis’s index as DeepSeek V4 Flash, a 284B mixture of experts that needs two H200s, at a tenth of the memory. The hardware profile is the useful part. About 54 GB of weights at BF16, which fits an H100, H200, or RTX Pro 6000. About 27 GB at FP8, which fits an L40S or a 48 GB card comfortably. Around 16 GB at 4-bit, which runs on a 24 GB consumer GPU with room for context. GGUF builds for llama.cpp and Ollama appeared within a day of release, and Unsloth had fine-tuning support the next day. If Qwen 4 27B follows the pattern, and the name says it will, the same three tiers apply: a 24 GB card for quantized local use, a 48 GB card for FP8, an 80 GB card for full precision. The [local Ollama and GGUF guide](https://www.yottalabs.ai/post/how-to-run-qwen-3-8-27b-locally-ollama-gguf-single-gpu-2026) and the [Unsloth fine-tuning guide](https://www.yottalabs.ai/post/how-to-fine-tune-qwen-3-8-27b-with-unsloth-2026) for the 3.8-27B are the closest thing to a Qwen 4 27B setup guide until the real one exists. ## The architecture question Here’s the one thing that could change the math. Alibaba released [Qwen 3.8-Flash-Next](https://www.yottalabs.ai/post/qwen-3-8-flash-next-specs-qwen-4-preview-2026) on August 28 as “an experimental preview of the architecture that will underpin Qwen4”: a sparse MoE with 125B total parameters and 6B active, plus a 51B-parameter n-gram memory table that trades computation for lookup, natively multimodal. That’s the Flash tier’s design. Whether the 27B inherits it is unknown, and it matters. A 27B dense model is 27B parameters in memory, full stop. A 27B model built the Flash-Next way could carry an n-gram table on top, and the “27B” in the name might count only the transformer. Flash-Next’s 125B plus 51B is the cautionary example: the headline number and the memory footprint are different numbers. Until the model card exists, plan for the dense case and keep headroom for the other. The active-parameter side cuts the other way. If Alibaba does bring lean activation to the 27B, it gets faster per token than the current dense 27B, which runs about 42 output tokens per second on hosted APIs and is the slowest of its peers. Speed is the current 27B’s weakest point; a Qwen 4 architecture would most plausibly fix it. ## Release date: what the cadence says No date. The strongest signal is how Qwen 3.8 rolled out: Max launched on August 3, and the 27B weights arrived on Hugging Face on August 13 to 14, ten days later, with the Max checkpoint around the same time. If Qwen 4 follows that shape, the 27B weights land a week or two after the flagship announcement, not on the same day. Alibaba said “very soon” for Qwen 4 on September 22; a public prediction market prices a Qwen 4 launch before November 1 at 74 percent. Read that as weeks, not a date, with the 27B trailing the headline by a little. ## How to prepare **Have the hardware you’d use for Qwen3.8-27B.** A 24 GB consumer card, a 48 GB workstation card, or an 80 GB datacenter GPU, depending on precision. If you don’t own one, [single-GPU pods by the hour](https://www.yottalabs.ai/pricing) cover the range, and the [production guide for the 27B](https://www.yottalabs.ai/post/how-to-run-qwen-3-8-in-production) is the serving setup that will carry over. **Run Qwen3.8-27B now.** Your prompts, your eval set, your quantization choice. The day the Qwen 4 27B card appears, the only question worth answering is “better or not on my workload,” and you can’t answer it without a baseline. **Read the license before you build on it.** The 3.8-27B was Apache 2.0, and Alibaba has said the Qwen 4 27B is open weights, but the 3.8 generation shipped three different licenses across its tiers (Apache 2.0 for the 27B, a custom license for Max, a community license for Flash-Next). Open weights and permissive weights are not the same thing. **Keep serving behind an OpenAI-compatible interface.** Qwen3.8-27B is live on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) at $0.375 in / $2.25 out per million tokens, alongside Qwen 3.8-Max, DeepSeek, GLM, and Kimi. When Qwen 4 27B lands there, the A/B is a model-string change. The [Best Chinese LLMs roundup](https://www.yottalabs.ai/post/best-chinese-llm-models-2026-deepseek-qwen-glm-kimi-compared) has the whole field the 27B will be judged against. ## Frequently asked questions **When is Qwen 4 27B coming out?** No date. Alibaba announced it on September 22 as part of the Qwen 4 family and said Qwen 4 is in training and coming “very soon.” In the 3.8 generation, the 27B weights arrived about ten days after the flagship launch. **Is Qwen 4 27B confirmed?** Yes, by name and role: Alibaba listed it at Apsara as the open-weights tier of Qwen 4 for local deployment. Nothing else about it is confirmed. **Will Qwen 4 27B be open source?** Alibaba said open weights. The license is unannounced; the current Qwen3.8-27B is Apache 2.0, and that’s the precedent, not a promise. **What hardware will Qwen 4 27B need?** Unknown, but if it’s a 27B dense model like its predecessor, plan on the same tiers: a 24 GB GPU at 4-bit, 48 GB at FP8, 80 GB at BF16. If it adopts the Flash-Next architecture with an n-gram memory table, the footprint could be larger than the name suggests. **Can I run Qwen 4 27B on Ollama or a 24 GB card?** Nothing to run yet. Qwen3.8-27B does both today, and the 27B line has had GGUF builds within a day of every release. **How is Qwen 4 27B different from Qwen 4 Flash?** Flash is an API tier built for low latency and high volume; Flash-Next is its architecture preview. The 27B is the tier you download and run yourself. Both are unreleased. **Is Qwen 4 27B on Yotta?** It doesn’t exist yet. Qwen3.8-27B is live on Yotta AI Gateway and runs on single-GPU pods; this post will be updated when Qwen 4 27B ships and again if it joins the catalog. ## Bottom line Qwen 4 27B is a confirmed name with an unconfirmed everything else. The current 27B tells you what to expect in role and hardware: the one Qwen 4 model built for a single GPU, most likely in the same 24 / 48 / 80 GB tiers, with an open question about whether the new architecture adds a memory table on top. The preparation is the same as for any model in this line: run the current one against your workload, have the card ready, and keep your serving one model string away from the swap. Qwen3.8-27B is on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) and on single-GPU pods today; the [Qwen 4 page](https://www.yottalabs.ai/post/qwen-4-release-date-what-is-known-how-to-prepare-2026) will have the family news first.